Model selection in time series by Deep Learning

Abstract

In this thesis, we will explore the use of deep learning techniques for model selection in time series. We compare the results from this with more traditional approaches for model selection, namely the Akaike and Bayesian information criterion. Specifically, we simulate data from AR(p), MA(q) and, ARMA(p,q) time series of three different lengths. Neural network models, such as fully connected, convolutional (CNN) and, Long Short-Term Memory (LSTM) models, are trained on this data to classify the true order of each sample. The accuracy of the Akaike and Bayesian information criterion and the accuracies of the neural network models in classifying the correct order are then compared. We found that deep learning models outperform or perform as well as the information criterion method in selecting the true order for each dataset.